Machine Learning

Machine learning in finance

• Fraud detection
• Credit risk analysis
• Return prediction
• Valuation
• Text analysis
• Time series forecasting

Models

• Linear
• Trees
• Neural networks
• Others

Regression vs Classification

• Regression means to predict a continuous variable (not necessarily linear regression).
• Classification is to predict a categorical variable. Binary or multiclass.

Train and Test

• Training means fitting a model (like linear regression).
• Objective is to make accurate predictions on new data.
• To assess performance, we have to check the model on “new data” (data not used in training).
• Split data into random train and test subsets. Train on training data. Test on test data.

Test criteria

• How do we decide if performance is good or bad?
• For continuous variables,
  • usually want to achieve a low sum of squared errors
  • equivalently, achieve a high \(R^2\).

\[R^2 = 1 - \frac{\sum (y_i - \hat y_i)^2}{\sum (y_i - \bar{y})^2}\]

• Categorical also based on prediction errors

Example data

• Download ml1.xlsx from the course website
• Upload it to Julius and ask Julius to read it and describe it.
• The data was created by generating 51 sets of 100 standard normals.
  • The first 50 sets are labeled x1, …, x50.
    
  • The last set was used as the noise to generate y1 as x1 + noise.
  • So, x2, …, x50 are irrelevant for y1.

Linear regression example

• Ask Julius to do a train-test split of the data with 20% of the data in the test set.
• Ask Julius to train a linear regression on the training data with x1, …, x50 as the features and y1 as the target.
• Ask Julius to compute the R-squared on the test data.
• Ask Julius to report the parameter estimates.

TBD